Ultramarine 1 Flash claims Fable-class knowledge work at $0.13 per task
- Ultramarine 1 Flash from Paper Instruments is said to match Claude Fable 5 on knowledge work in finance, law, consulting, and operations, at 60x lower cost.
- The model is post-trained with large-scale reinforcement learning on GLM-5.3 Flash, and it generates 6.5% fewer output tokens than that base across six benchmarks.
- On the company's internal PI Knowledge Work benchmark it scores 59.3 against 57.8 for Claude Fable 5, and the company says that benchmark correlates with the Artificial Analysis Intelligence Index (r = 0.81).
- The gap closes or reverses on outside tests: it scores 57.0 on APEX-Agents 1.1 against 63.6 for Claude Fable 5, and it ranks second on AutomationBench behind DeepSeek V4.1 Flash.
- The API costs $0.15 per million input tokens and $0.50 per million output tokens, and the company puts a task at $0.13 versus $8.19 for Claude Fable 5 (max), with model weights on GitHub promised as coming soon.
Hacker News opinions
Why's the cost per task chart missing 6.1 Sol? Feels like an easy one to leave off.
Ultramarine for the doc work and Jev for the decisions, if I read it right. Sounds useful in an office. I'd be more excited if I worked at a big corp.
Nice that they say it's built on GLM-5.3 Flash. I can't judge the post-training from outside, but at least you know what's under the hood.
They never said frontier model, they said frontier-level performance. Different thing.
People mostly mean the cost/performance Pareto frontier these days. Best score you can get at that price.
Agree with the pushback though. Frontier means pushing intelligence, like Fable or Astra. For knowledge work the model fit matters more than price.
Doesn't look like coding was the focus. DeepSWE is 61.5 here vs 63.4 for GLM 5.3 Flash, so it's a step down. I wouldn't reach for it for code.
The Results section only compares against models it beats on every metric. Then the cost chart brings back Fable 5, GPT-6 Astra, and Opus 5.5. Just own your weak spots.
Same thing happened to GLM and Qwen, dozens of fine tunes with cute names. I'd like to see the Thomson Reuters model run on these benchmarks too, it's smaller and genuinely useful for document analysis.
Bullish on distilled models like this, especially if it takes some wind out of the Anthropic and OpenAI IPO plans.
For bookkeeping like invoice sorting I'd probably just write a python script before throwing an LLM at it.